01 / The short version
What happened
PyTorch 2.15 adds FP8 tensor support on MPS for Apple silicon Macs running macOS 15 or later. The release accelerates language model decoding through dedicated Metal kernels for linear layers, mixture-of-experts grouped matmul, and softmax operations. It also introduces complex tensor operations, matrix inversion and SVD improvements, and new Metal paths for audio, video, and training workloads. The publisher reports a 1.23x geometric-mean speedup across 96 BF16 layer shapes and approximately 1.3x higher token throughput for MoE models on an M3 Ultra with torch.compile. Constant padding changes reduce peak driver memory in Wan2.2 VAE decode from about 270 GiB to 8 GiB. Users can now select MPS streams directly from Python.
See the exact references02 / Key takeaways
What you need to know
- 01
FP8 tensors are now supported on MPS with torch.float8_e4m3fn.
- 02
Dedicated Metal kernels speed up dense linear, MoE grouped matmul, and softmax operations.
- 03
Constant padding on Metal reduces peak driver memory from about 270 GiB to 8 GiB in Wan2.2 VAE decode.
Keep in perspective
What to watch for
Performance claims are attributed to the publisher and have not been independently validated.
Go to the source
Exact references
These are the original pages used for this brief. Publisher claims are not independent evaluations.
01Primary source · Hugging FaceRead the original announcementhttps://huggingface.co/blog/mps-pytorch-2-15AI-generated from the linked source. It can miss context; verify consequential details in the original. How the radar works